JP2014085795A - Learning image collection device, learning device and object detection device - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a learning image collection device easily collecting images in environment different from that of an image when a detector has learned, a learning device using the collected images, and an object detection device.SOLUTION: A learning image collection device comprises an acquisition part, an extraction part, a calculation part and a selection part. The acquisition part acquires an image including an object. The extraction part extracts a plurality of candidate areas being a candidate of the object from the image. The calculation part calculates either one of first similarity between the candidate areas and a prescribed area, second similarity between sizes of the candidate areas and the prescribed object, or third similarity of each of the plurality of candidate areas. The selection part selects a candidate area as an object area including the object when either one of the first, second, or third similarity is larger than a prescribed threshold.

Description

  Embodiments described herein relate generally to a learning image collection device, a learning device, and an object detection device.

  A technique related to a recognition apparatus that can collect learning patterns more effectively by identifying whether or not the object is an object from the captured image and tracking the object is disclosed. In this technique, in order to collect efficiently, images of an object in the case of misidentification are collected.

  However, if it cannot be identified in advance whether the object is an object, it cannot be traced and a new learning image cannot be collected. On the other hand, even when the detector is learned using a large amount of learning data, the detection rate is lowered in an image in an environment different from that at the time of learning.

JP2007-310480

  The problem to be solved by the present invention is to provide a learning image collection device that easily collects an image having a different environment from the learning image of the detector, a learning device using the collected image, and an object detection device. is there.

  The learning image collection device of the embodiment includes an acquisition unit, an extraction unit, a calculation unit, and a selection unit. An acquisition part acquires the image containing a target object. The extraction unit extracts a plurality of candidate regions that are candidates for the object from the image. The calculation unit is any one of a first similarity between the candidate area and the predetermined area, a second similarity with the predetermined size of the object, and a third similarity with each of the plurality of candidate areas. Calculate similarity. The selection unit selects an object region including the object when any one of the first similarity, the second similarity, and the third similarity is larger than a predetermined threshold.

  The learning device according to the embodiment includes an acquisition unit, an extraction unit, a calculation unit, a selection unit, and a learning unit.

An acquisition part acquires the image containing a target object. The extraction unit extracts a plurality of candidate regions that are candidates for the object from the image. The calculation unit is any one of a first similarity between the candidate area and the predetermined area, a second similarity with the predetermined size of the object, and a third similarity with each of the plurality of candidate areas. Calculate similarity. The selection unit selects an object region including the object when any one of the first similarity, the second similarity, and the third similarity is larger than a predetermined threshold. The learning unit learns a discriminator for identifying an object using the object area as teacher data.

  The target object detection apparatus according to the embodiment includes an imaging unit, an extraction unit, a calculation unit, a selection unit, a learning unit, and a detection unit. The imaging unit captures an image including the object. The extraction unit extracts a plurality of candidate regions that are candidates for the object from the image. The calculation unit is any one of a first similarity between the candidate area and the predetermined area, a second similarity with the predetermined size of the object, and a third similarity with each of the plurality of candidate areas. Calculate similarity. The selection unit selects an object region including the object when any one of the first similarity, the second similarity, and the third similarity is larger than a predetermined threshold. The learning unit learns a discriminator that identifies an object using the object region as teacher data. The detection unit detects the object from the image.

The lineblock diagram showing the example of the learning image collection device of a 1st embodiment. The flowchart which shows the example of the learning image collection device of 1st Embodiment. Explanatory drawing which shows an example at the time of setting a person as a detection target. Explanatory drawing which shows an example of an image required for the learning which considered environmental conditions. The figure which shows an example of the screen of the modification of 1st Embodiment. The figure which shows an example of the screen of the modification of 1st Embodiment. The figure which shows an example of the screen of the modification of 1st Embodiment. The block diagram which shows the example of the learning apparatus of 2nd Embodiment. The block diagram which shows the example of the target object detection apparatus of 2nd Embodiment. The flowchart which shows the example of the learning apparatus of 2nd Embodiment. The block diagram which shows the example of the target object detection apparatus of 3rd Embodiment. The flowchart which shows the example of the target object detection apparatus of 3rd Embodiment. The block diagram which shows the modification of 3rd Embodiment. The block diagram which shows the modification of 3rd Embodiment.

  Hereinafter, embodiments will be described in detail with reference to the accompanying drawings.

(First embodiment)
FIG. 1 is a configuration diagram illustrating an example of a learning image collection device 1 according to the first embodiment. As shown in FIG. 1, an acquisition unit 10, an extraction unit 11, a calculation unit 12, and a selection unit 13 are provided. The learning image collection device may be realized by causing a processing device such as a CPU (Central Processing Unit) to execute a program, that is, by software, or by hardware such as an IC (Integrated Circuit). It may be realized by using software and hardware together. The image acquired by the acquisition unit 10 may be an image stored in a storage device.

The storage device is, for example, at least one of magnetic, optical, and electrical storage devices such as a hard disk drive (HDD), a solid state drive (SSD), a read only memory (ROM), and a memory card. Can be realized.

  FIG. 2 is a flowchart showing an example of a flow of processing of the learning image collection device 1 according to the first embodiment.

First, the learning image collection device 1 acquires an image captured by the imaging device or an image stored in the storage device (step S11). Next, the extraction unit extracts a region that can be a candidate for the target object as a candidate region (step S12). The extraction of the candidate area may be performed using inter-frame difference processing or background difference processing. For example, using the method of Wallflower: Principles and Practice of Background Maintenance (Kentaro Toyama, John Krumm, Barry Brumitt, Brian Meyers, Proceedings of the 7th IEEE International Conference on Computer Vision, ICCV 1999, pp255-261, September, 1999) Good candidate areas include, for example, fluctuating backgrounds such as people, vehicles, swaying trees, doors that open and close.

  FIG. 3 is a diagram showing an example when a person is a detection target. As described above, in the inter-frame difference and background difference processing, it is not possible to select only the target to be collected as learning data.

  FIG. 4 shows an example of an image necessary for learning in consideration of environmental conditions. In the object detection process, it is ideal that detection is possible even when the environmental conditions are different from those during learning (X1). However, in practice, the object is often detected in an environment where the sunlight conditions (light and its direction) are different. Until now, detection has been difficult if the environmental conditions (ambient conditions) differ from those during learning (X2). By using difference detection such as inter-frame difference and background difference processing, learning images for object detection processing can be collected from images with different environmental conditions. For example, tree shaking and door opening / closing are detected (X3), and it is not always possible to collect only images suitable for learning.

  In the case of detecting a person, it is desirable to obtain a learning image for a single person when the environmental conditions are different. This is because when the detection target is “person”, if a single person can be detected, a plurality of persons can be detected. That is, since the proposed method can select a single person from the acquired images with different environmental conditions (X4), by additionally learning this as teacher data, an apparatus capable of detecting an object even when the environmental conditions are different It becomes possible to provide (X5).

  In order to efficiently collect objects with different environmental conditions, (a) compare the aspect ratios of the rectangles in the images with different environments as in the case of FIG. 3, and (b) change the positions on the images. Use conditions. Details of the conditions will be described with reference to the steps in FIG.

  Returning to FIG. When the extraction unit extracts a plurality of candidate areas, the user selects one image necessary for learning (step S13). When detecting a person, only an area including the person is selected. For example, when it is desired to detect a vehicle, only an area including the vehicle needs to be selected. When it is desired to detect an object other than a person or a vehicle, the same processing may be performed on the object.

  The calculating unit 12 compares the target candidate area with the selected or designated target area to calculate the similarity S1 (step S14). Here, the similarity is calculated using, for example, normalized correlation processing between the target candidate area and the designated target area.

  Further, the calculation unit 12 calculates the similarity S2 between the size information of the target candidate area and the size information designated in advance (step S15). Here, as the size information, for example, the size and the aspect ratio of the target in the vertical and horizontal directions are used.

The similarity S2 indicates that the horizontal length of the candidate area A is width_A, the vertical length is height_A, the predetermined horizontal length of the target area B is width_B, and the vertical length is height_B. Similarity is based on area, for example (Formula 1)
1 / {| (width_A * height_A)-(width_B * height_B) | / (width_B * height_B)}
And calculate. Based on the aspect ratio (Equation 2)
1 / {| (width_A / height_A)-(width_B / height_B) | / (width_B / height_B)}
May be calculated. Further, the product of Equation 1 and Equation 2 described above may be used.

  A candidate area having a small calculated value of S2 means that the difference from the size information designated in advance is large. For example, it can be estimated that a candidate region having a large value of S2 is highly likely to be a person. When a person is designated as a detection target, a candidate area with a small value of similarity S2 corresponds to a vehicle area or the like. At the time of image collection, it is handled by a method of excluding candidate areas having a small value of similarity S2. This is preferable because the accuracy of the detection device is improved.

  For example, when a person is designated as a detection target, a region in which a plurality of persons overlap and are extracted as one group is not preferable as learning data. When the aspect ratio is used, the value of the similarity S2 becomes small because one group in which a plurality of persons overlap has a greatly different aspect ratio. Since these can be excluded from the learning data, the accuracy of the detection device is improved, which is preferable.

In addition, the calculation unit 12 compares the obtained plurality of target candidate regions and calculates the similarity S3 (step S16). Here, the similarity may be calculated using, for example, a general normalized correlation process. For example, when there is a candidate area X and other N candidate areas X_i (i = 1 to N),
(Formula 3)
S3 = 1 / N * Σi = 1 to N (Sim (X, X_i)) (where Sim (A, B) is the normalized correlation value of A and B)
And it is sufficient. When the value of the similarity S3 is small, it means that the pattern is greatly different from other compared candidate areas. For example, there is a high possibility that the background or the like is suddenly extracted. In this case, it is preferable to exclude it from the learning data because the accuracy of the detection device is improved.

  Furthermore, the calculation unit 12 calculates the degree of difference S4 between the positions of the plurality of target candidate regions (step S17). Here, the degree of difference in position is expressed by the following equation (Equation 4) when there is a certain candidate region X and M candidate regions X_i (i = 1 to M) having sufficiently high similarity to the candidate region X. it can.

(Formula 4)
S4 = 1 / M * Σi = 1 to M (Dist (X, X_i)) (where Dist (A, B) is the Euclidean distance between A and B). It can be estimated that the pattern appears only in. For example, it is a case where a background area different from a person who is a target to be detected is repeatedly extracted, such as opening and closing of an automatic door. For example, it can be estimated that a candidate region having a large value of S4 is highly likely to be a person. In addition, when collecting learning images, it is preferable to use a method of excluding candidate regions having a small value of similarity S4 because the accuracy of the detection device is improved.

  The selection unit 13 determines whether the similarity or dissimilarity (S1 to S4) calculated by the calculation unit 12 is greater than a predetermined threshold (step S18), and selects a learning image from the candidate area. The selection from the candidate area is either the similarity or the difference (S1 to S4), or the image area is selected based on the result of two or more similarities (step S19).

  For example, weights w1 to w4 may be set for the calculated values of S1 to S4, and the following weighting evaluation may be performed.

(Formula 5)
(w1 * S1) + (w2 * S2) + (w3 * S3) + (w4 * S4)
The selection unit 13 selects a region where the evaluation value is equal to or greater than a predetermined threshold as a learning image. Those having an evaluation value smaller than a predetermined threshold value are preferably excluded because they are excluded from the learning data and the accuracy of the detection target of the detection device is improved.

  Further, the calculation unit 12 calculates the movement amount S5 of the candidate area (step S21). For example, the target candidate may be tracked over a plurality of images in time series order, and the movement amount may be calculated. For example, the following method may be used for the tracking process.

Dorin Comaniciu, Visvanathan Ramesh, Peter Meer: Real-Time Tracking of Non-Rigid Objects using Mean Shift, pp.142--149, Proceedings of the 2000 IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR 2000)
It is determined whether or not the calculated movement amount S5 is greater than a predetermined threshold (step S22), and a learning image is selected from the candidate area. The selection from the candidate area is either the similarity or the difference (S1 to S5), or the image area is selected based on the result of two or more similarities (step S19).
For example, similarly to the above-described Expression 5, the weights w1 to w5 may be set for the calculated values of S1 to S5, and the following weighting evaluation may be performed.

(Formula 5)
(w1 * S1) + (w2 * S2) + (w3 * S3) + (w4 * S4) + (w5 * S5)
The selection unit 13 selects a region where the evaluation value is equal to or greater than a predetermined threshold as a learning image. Those having an evaluation value smaller than a predetermined threshold are excluded from the learning data. In this case, the accuracy of the detection target of the detection device is improved, which is preferable.

  The selected candidate area may be presented on a display device or the like so that the user can select it (step S20).

  As described above, according to the learning image collecting apparatus 1 according to the first embodiment, it is possible to easily collect images having different environments from the images when the detector is learned. In particular, when trying to detect an object from an image of an environment different from the time of learning of the detector, tracking or additional learning may not be possible, but by using the detection target image collecting apparatus according to the present embodiment, It is possible to obtain a detection target image suitable for detector learning.

(Modification)
The detection target image collection apparatus according to the modification of the first embodiment may include a manual selection unit instead of the selection unit 13 in the first embodiment. First, from step S1 to step S18, S21, and S22 in the first embodiment, the same processing is performed in the present modification.

  In step S19, the manual selection unit arranges the image regions based on the calculation results of any one or two or more of the similarity S1 to the similarity S3, the difference S4 and S5. Here, it is preferable to perform weighting evaluation in the same manner as in step S19 in the first embodiment, and arrange them in descending order of the evaluation value.

  FIG. 5 shows an example in the case where the evaluation values are arranged in descending order. The manual selection unit displays the aligned images on the display unit. The user determines which image to select as the learning image, and the detection target image collection device accepts the user's specification.

    Here, the image area candidates for learning presented to the user are arranged in the order in which they are estimated to be more appropriate as images for learning. For this reason, the user can preferentially check an image area having a high evaluation value. In FIG. 5, the evaluation values are arranged higher as they are on the left. Therefore, the user may confirm the candidate area from the left. Since the image region candidate on the left side may be checked with priority, the learning image can be easily selected. By such user confirmation, it is possible to prevent an inappropriate image region from being selected. In addition, since the user can exclude the image area having a low evaluation value by simple confirmation, it is possible to reduce the trouble of confirmation.

  Further, it may be displayed including the periphery of the image to be selected. By displaying the image including the periphery of the image to be selected, it is possible to confirm and determine the periphery of the target region to be confirmed. For example, FIG. 6 shows an example in which an image area to be confirmed is superimposed on the entire image. When a person is a target, a frame around the image may be displayed, or an evaluation value may be reflected in the color and thickness of the frame, the type of line, and the like. For example, the color of the frame is displayed in red, and the darker color has a larger evaluation value, the evaluation value is high for red, and the evaluation value is low for blue. Moreover, the evaluation value may be larger as the thickness of the frame is larger, or the evaluation value may be smaller by changing the type of the frame line to a chain line. For example, when the evaluation value is the highest, a symbol such as an arrow may be displayed. When displaying and selecting surrounding images, it is easy for the user to understand what part of the image has been selected.

On the other hand, since the magnitudes of the evaluation values are difficult to understand, mode switching such as rearranging in order of evaluation values may be presented separately. FIG. 7 is a diagram showing an example in the case where FIGS. 5 and 6 are combined. Arranging the evaluation values in descending order is preferable because it improves the visibility of the user and improves the visibility of the user, and also displays the actual state of the image together, making it easier to grasp the overall state. These may be "put only the candidates that exist on the screen above",
Mainly on the lower screen, a mode may be presented in which “the lower is arranged in the order of evaluation values in all videos and the corresponding whole image is displayed on the upper side when the lower candidate is designated”.

(Second Embodiment)
FIG. 8 is a block diagram showing a learning device 2 according to the second embodiment of the present invention. The learning device according to the present embodiment includes an acquisition unit 10, an extraction unit 11, a calculation unit 12, a selection unit 13, and a learning unit 24. The difference from the first embodiment is that a learning unit 24 is provided.

  Moreover, FIG. 9 is the target object detection apparatus 3 using the learning apparatus concerning the 2nd Embodiment of this invention. The first embodiment is different from the first embodiment in that the imaging unit 30, the learning unit 24, and the detection unit 35 are provided. 8 is different from FIG. 8 in that an imaging unit 30 is provided.

  Hereinafter, the same configurations and processes as those of the first embodiment are denoted by the same reference numerals, and description thereof is omitted.

  Next, the operation of the object detection apparatus according to the second embodiment of the present invention will be described with reference to FIG. FIG. 10 is a flowchart showing the operation of the object detection apparatus according to the second embodiment of the present invention.

  First, an image is captured by the imaging unit 30. The acquired image is acquired by the acquisition unit 10 in the case of the learning apparatus of FIG. Next, from S12 in the first embodiment to steps S18, S21 and S22, the same processing is performed in the second embodiment.

  In step S23, detection target images suitable for detector learning are selected, and the learning unit 24 calculates and learns information necessary for detection using these images (step S24). For example, G. Cauwenberghs and T. Poggio, "Incremental and Decremental Support Vector Machine Learning," in Adv. Neural Information Processing Systems pp.409-415 (NIPS * 2000), Incremental and Decremental Support Vector Machine Learning Additional learning may be performed. The result calculated by learning is information that is commonly called a dictionary in the field of image recognition and serves as a clue to detection. Based on the calculation result, the detection unit 35 detects the object of the designated image (step S25).

Here, in step S24, another detector that operates in parallel with the existing detector may be newly learned and used as an additional detector. Then, in step S25, a result common to the result of the existing detector and the result of the newly added detector is detected. Or you may use both the result of the existing detector, and the newly added detector as a detection result. For example, when the detection parameters used for the existing detector and the newly added detector are different, the detection performance can be improved by using the result of comparing both detection results. According to the learning device 2 according to the embodiment, it is possible to easily collect images having different environments from the images when the detector is learned. Moreover, according to the target object detection apparatus 3 which concerns on 2nd Embodiment, the image from which an environment differs can be collected and the detection performance of a target object can be improved. In particular, when an object is to be detected from an image in an environment different from that at the time of learning of the detector, tracking or additional learning may not be possible. However, by using the learning device according to the present embodiment, Since a detection target image suitable for learning can be obtained, the detection performance of the target can be improved.

(Third embodiment)
FIG. 11 is a configuration diagram illustrating an example of the object detection device 3 according to the third embodiment. The second embodiment is different from the second embodiment in that a data distribution determination unit 46 (46A and 46B) is provided. Hereinafter, the same configurations and processes as those in the first embodiment and the second embodiment are denoted by the same reference numerals, and the description thereof is omitted.

  Next, the operation of the object detection apparatus according to the third embodiment of the present invention will be described using FIG. 11 and FIG. FIG. 12 is a flowchart showing the operation of the object detection apparatus according to the third embodiment of the present invention.

  Hereinafter, a case where the learning image collection device 1A collects images and distributes the image information to the object detection device including the learning image collection device 1B will be described. First, from S11 to steps S18, S21, and S22 in the first embodiment, the same processing is performed in this modification.

  In step S23, a detection target image suitable for the learning of the detector is selected, and the data distribution determination unit 46 determines a distribution destination of the image information of the selected image (step S26). At this time, for example, an object detection device installed at a geographically close position may be selected as the distribution destination. Alternatively, the image information distributed in the past from the distribution destination candidates is compared with the information on the detection target image obtained in step S23, and the object detection apparatus having the image information in the range in which the similarity is specified in advance is selected as the distribution destination. Can be specified as

  The learning means calculates information necessary for detection using the obtained image information (step S27). This calculation may be performed in the same manner as step S25 in the second embodiment. Based on the calculation result, the detection unit 35 performs a detection process (step S28).

  Through the above processing, not only the detection target image collected by the learning image collection device 1A but also the detection target image collected by another learning image collection device 1B can be obtained. There may be two or more object detection devices.

  For example, in step S26, an object detection apparatus having image information with a high degree of similarity is designated as a distribution destination. An object detection apparatus having image information with a high degree of similarity is an object detection apparatus having image information with similar depression angles, illuminances, and the like. In this case, the image of the object detection device that captures an image with a large depression angle is distributed to another object detection device that also captures an image with a large depression angle, and the object detection device that captures an image with low illuminance This image is also distributed to another object detection device that captures an image with low illuminance. As a result, a larger amount of detection target images can be collected and learned, and higher-performance target detection can be performed.

  Conversely, when an object detection device having image information with low similarity is designated as the delivery destination in step S26, for example, the image of the object detection device that captures an image with a large depression angle is an image with a small depression angle. The image of the object detection device that captures an image with low illuminance is distributed to another object detection device that captures an image with high illuminance. As a result, more diverse detection target images can be collected and learned, and more stable target detection can be performed.

(Modification 1)
In FIG. 11, the object detection devices are configured to distribute image information to each other, but as illustrated in FIG. 13, the data distribution determination unit 46 may be independent. In this case, for example, the data distribution determination unit exists in the central server, and the image information collected by the detection target image collection device 1A or 1B is once transmitted to the data distribution determination unit 46 of the central server. The data distribution determining unit 46 determines the distribution destination of the image information and distributes it to each target object detection device. There may be two or more object detection devices.

(Modification 2)
In FIG. 11, the object detection devices are configured to distribute image information to each other. However, as illustrated in FIG. 14, the selection unit, the learning unit, and the data distribution unit 46 may be provided in the central server. In this case, the selection unit 53 compares and selects image information obtained from the plurality of learning image collection devices (1A, 1B). For example, an image of a learning image collection device installed at a geographically close position is selected. Alternatively, image information obtained from a plurality of learning image collection devices is compared, and an image of the learning image collection device having image information in a range in which the degree of similarity is designated in advance is selected.

  The learning unit 35 calculates information necessary for detection using the obtained image information. The data distribution determination unit 46 of the central server distributes the calculation result to the detection unit 35A or 35B determined as the distribution destination, and the detection unit 35A or 35B performs detection based on the distributed information.

(Hardware configuration)
The learning image collection device, the learning device, and the object detection device of the above embodiment are a control device such as a CPU (Central Processing Unit), a storage device such as a ROM and a RAM, an external storage device such as an HDD and an SSD, and a display such as a display. The apparatus includes an input device such as a mouse and a keyboard, an imaging device such as a camera, and the like, and can be realized with a hardware configuration using a normal computer.

  The program executed by the apparatus of the above embodiment is provided by being incorporated in advance in a ROM or the like.

  The program executed by the apparatus of the above embodiment can be read by a computer such as a CD-ROM, a CD-R, a memory card, a DVD, and a flexible disk (FD) in an installable or executable file. It may be provided by being stored in a different storage medium.

  The program executed by the apparatus of the above embodiment may be provided by storing it on a computer connected to a network such as the Internet and downloading it via the network. The program executed by the apparatus of the above embodiment may be provided or distributed via a network such as the Internet.

  The program executed by the apparatus of the above embodiment has a module configuration for realizing the above-described units on a computer. As actual hardware, for example, the control device reads out a program from an external storage device to the storage device and executes the program, whereby the above-described units are realized on a computer.

  Note that the present invention is not limited to the above-described embodiment as it is, and can be embodied by modifying the constituent elements without departing from the scope of the invention in the implementation stage. In addition, various inventions can be formed by appropriately combining a plurality of components disclosed in the embodiment. For example, some components may be deleted from all the components shown in the embodiment. Furthermore, constituent elements over different embodiments may be appropriately combined.

  For example, as long as each step in the flowchart of the above embodiment is not contrary to its nature, the execution order may be changed, a plurality of steps may be performed simultaneously, or may be performed in a different order for each execution.

  As described above, according to the above-described embodiment, it is possible to easily collect images having different environments from the images at the time of learning of the detector, and it is possible to improve the detection performance of the object by learning.

DESCRIPTION OF SYMBOLS 1 ... Learning image collection apparatus, 2 ... Learning apparatus, 3 ... Object detection apparatus, 10 ... Acquisition part, 11 ... Extraction part, 12 ... Calculation part, 13, 53. ..Selection unit, 24 ... learning unit, 30 ... imaging unit, 35 ... detection unit, 46 ... data distribution determination unit

Claims (14)

An acquisition unit that acquires an image including an object, an extraction unit that extracts a plurality of candidate areas that are candidates for the object from the image, a first similarity between the candidate area and the predetermined area, and the predetermined object A calculating unit that calculates any one of the second similarity with the size of the first similarity and the third similarity with each of the plurality of candidate regions;
A selection unit configured to select an object region including the object when any one of the first similarity, the second similarity, and the third similarity is greater than a predetermined threshold. Learning image collection device. The acquisition unit acquires a plurality of time-series images including the object,
The extraction unit further extracts a plurality of candidate region positions,
The calculation unit further calculates a change in position of the plurality of candidate regions in time series as a fourth similarity,
The selection unit selects the candidate region as the target region when any one of the first similarity, the second similarity, the third similarity, and the fourth similarity is greater than a predetermined threshold. The learning image collecting apparatus according to claim 1, wherein: The calculation unit further calculates the amount of movement of the candidate area between time series images using the similarity of each of the candidate areas of the time series image,
The learning image collecting apparatus according to claim 2, wherein the selection unit sets the candidate region as the target region when the movement amount is larger than a predetermined threshold. The object is a person,
The learning image collecting apparatus according to claim 1, wherein the size used for the second similarity is an aspect ratio of the candidate area. The selection unit selects the object region using the second similarity and the fourth similarity,
The candidate area is excluded from the object area when the fourth similarity is smaller than a predetermined threshold among the candidate areas where the second similarity is larger than a predetermined threshold. 4. The learning image collecting apparatus according to 4. The selection unit selects the object region using the second similarity and the movement amount,
The candidate area is excluded from the object area when the movement amount is smaller than a predetermined site among the candidate areas having the second similarity greater than a predetermined threshold. Learning image collection device. An acquisition unit that acquires an image including an object, an extraction unit that extracts a plurality of candidate areas that are candidates for the object from the image, a first similarity between the candidate area and the predetermined area, and the predetermined object A calculating unit that calculates any one of the second similarity with the size of the first similarity and the third similarity with each of the plurality of candidate regions;
A selection unit that selects an object region including the object when any one of the first similarity, the second similarity, and the third similarity is greater than a predetermined threshold;
A learning apparatus comprising: a learning unit that learns a discriminator that identifies an object using the object region as teacher data. The object is a person,
The learning apparatus according to claim 7, wherein the size used for the second similarity is an aspect ratio of the candidate area. The selection unit selects the object region using the second similarity and the fourth similarity,
Of the candidate regions with the second similarity greater than a predetermined threshold, when the fourth similarity is smaller than a predetermined threshold, the candidate region is excluded from the target region to remove the target region The learning device according to claim 8, wherein the learning device is selected. The selection unit selects the object region using the second similarity and the movement amount,
The candidate area is selected by excluding the candidate area from the object area when the movement amount is smaller than a predetermined site among the candidate areas having the second similarity greater than a predetermined threshold. The learning device according to claim 8. An imaging unit that captures an image including an object;
An extraction unit that extracts a plurality of candidate areas that are candidates for the object from the image; a first similarity between the candidate area and the predetermined area; a second similarity with a predetermined size of the object; A calculation unit that calculates any one of the third similarities with each of the plurality of candidate areas;
A selection unit that selects a target region including the target when any one of the first similarity, the second similarity, and the third similarity is greater than a predetermined threshold value. An object detection apparatus comprising: a learning unit that performs learning of a classifier that identifies an object; and a detection unit that detects the object from the image. The object is a person,
The object detection apparatus according to claim 11, wherein the size used for the second similarity is an aspect ratio of the candidate area. The selection unit selects the object region using the second similarity and the fourth similarity,
The candidate area is excluded from the object area when the fourth similarity is smaller than a predetermined threshold among the candidate areas where the second similarity is larger than a predetermined threshold. 12. The learning device according to 12. The selection unit selects the object region using the second similarity and the movement amount,
13. The learning according to claim 12, wherein, among the candidate areas having the second similarity greater than a predetermined threshold, the candidate area is excluded from the object area when the movement amount is smaller than a predetermined site. apparatus.